基于三种高分辨率影像样本迁移学习的城市水体提取研究

水体提取是遥感监测城市水环境必不可少的步骤，提取城市中的细小水体目前已成为遥感影像深度学习领域的热点。但是，深度学习需要大量的样本数据集作为输入，而且不同空间分辨率影像往往需要构建不同的样本集。如果影像的空间分辨率差异不大，可以先采用分辨率较低的影像样本训练模型，并加入少量的较高分辨率样本再次训练模型，这种模型可以保证精度和节约时间。研究选用了U-net图像分割模型，针对3种不同空间分辨率——分别为0.5 m、0.8 m和2 m的影像进行样本迁移学习。发现2 m到0.8 m、2 m到0.5 m、0.8 m到0.5 m 3种迁移学习后，提取水体结果对应评价指标F1-score、MIoU、Kappa都在0.80以上。在分辨率差异不大的前提下，这种从较低分辨率样本迁移到较高分辨率影像提取城市水体的方法基本可行，结果精度较好，适用于缺水型城市的水体提取。

Research on Urban Water Body Extraction based on Transfer Learning of Three High-resolution Image Datasets

Water extraction is an essential step for rare earth monitoring of urban water environment. Extraction of small water bodies in the city has now become a hot depth study in the field of remote sensing images. However， deep learning requires a large number of sample datasets as input， and images with different spatial resolutions often need to construct different datasets. If the spatial resolution of the images is not much different， the sample transfer learning model can be used to ensure accuracy and save time. In this paper， the U-Net image segmentation model is selected to perform sample transfer learning for images with three different spatial resolutions—0.5 m， 0.8 m and 2 m respectively. It is found that after three migration learning of 2 meters to 0.8 meters， 2 meters to 0.5 meters， and 0.8 meters to 0.5 meters， the corresponding evaluation indexes F1-score， MioU and Kappa of the extracted water body are all above 0.80. Under the premise of little difference in resolution， this method of extracting urban water bodies from lower-resolution samples to higher-resolution images is basically feasible， and the accuracy of the results is better. It is suitable for water extraction in water-deficient cities.